1 / 20

Machine Learning: Learning finite state environments

Machine Learning: Learning finite state environments. Avrim Blum 15-451 lecture 12/09/04. Machine Learning. A big topic in Computer Science. We’d like programs that learn with experience. Because it’s hard to program up complicated things by hand.

mmarlowe
Download Presentation

Machine Learning: Learning finite state environments

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Machine Learning: Learning finite state environments Avrim Blum 15-451 lecture 12/09/04

  2. Machine Learning A big topic in Computer Science. We’d like programs that learn with experience. • Because it’s hard to program up complicated things by hand. • Want software that personalizes itself to users needs. • Because it’s a necessary part of anything that is going to be really intelligent.

  3. Pomerleau NHAA Schneiderman Kanade What ML can do • Learn to steer a car. • Learn to read handwriting, recognize speech, detect faces. • Learn to play backgammon (best in world). • Identify patterns in databases. Generally, program structure developed by hand. Learning used to set (lots of) parameters. ML as programmer’s assistant.

  4. More conceptually... • Can we use CS perspective to help us understand what learning is? • Think about learning as a computational task just like multiplying? • How does a baby learn to figure out its environment? To figure out the effect of its actions? • Lots of parts to all this. Today: one problem that captures some small piece of it.

  5. Imagine... • Say we are a baby trying to figure out the effects our actions have on our environment... • Sometimes actions have effects we can notice right away, sometimes effects are more long-term.

  6. A model: learning a finite state environment • Let’s model the world as a DFA. We perform actions, we get observations. • Our actions can also change the state of the world. # states is finite. start 0,1 1 Actions 0 and 1. Observations white or purple. 0 0 1

  7. Learning a DFA Another way to put it: • We have a box with buttons and lights. • Can press the buttons, observe the lights. • lights = f(current state) • next state = g(button, prev state) • Goal: learn predictive model of device.

  8. Let’s look at an easier problem first: state = observation. space S. Learning DFAs This seems really hard. Can’t tell for sure when world state has changed.ample space S.

  9. An example w/o hidden state 2 actions: a, b. • Generic algorithm for lights=state: • Build a model. • While not done, find an unexplored edge and take it. • Now, let’s try the harder problem!

  10. Some examples Example #1(3 states) Example #2(3 states)

  11. Our model: a,b Real world: a a b b a b a a b b a b Can we design a procedure to do this in general? One problem: what if we always see the same thing? How do we know there isn’t something else out there? Called “combination-lock automaton”

  12. Real world: a a b b a b a a b b a b Can we design a procedure to do this in general? Called “combination-lock automaton” This is a serious problem. It means we can’t hope to efficiently come up with an exact model of the world from just our own experimentation.

  13. How to get around this? • Assume we can propose model and get counterexample. • Alternatively, goal is to be predictive. Any time we make a mistake, we think and perform experiments. • Goal is not to have to do this too many times. For our algorithm, total # mistakes will be at most # states.

  14. Today: a really cool algorithm by Dana Angluin (with extensions by R.Rivest & R.Schapire) • To simplify things, let’s assume we have a RESET button. • If time, we’ll see how to get rid of that.

  15. The problem (recap) b b • We have a DFA: a a > a b • observation = f(current state) • next state = g(button, prev state) • Can feed in sequence of actions, get observations. Then resets to start. • Can also propose/field-test model. Get counterexample.

  16. l a b b l a b aa ab ba bb a a > a b Key Idea Key idea is to represent the DFA using a state/experiment table. experiments states trans-itions Either aa=b or else aa is a totally new state and we need another expt to distinguish.

  17. l a l a b aa ab ba bb Key Idea Key idea is to represent the DFA using a state/experiment table. experiments Guarantee will be: either model is correct, or else the world has > n states. In that case, need way of using counterexs to add new state to model. states trans-itions

  18. The algorithm We’ll do it by example... a a a b b > a

  19. Algorithm (formally) go to 1.

  20. Summary / Related problems • All states looks distinct: easy. • Not all look distinct: • can do with counterex. • All distinct but probabilistic transitions? • Markov Decision Process(MDP) / Reinforcement Learning. • Usual goal: maximize discounted reward (like probabilistic shortest path). DP-based algs. • Not all distinct & probabilistic transitions? • POMDP. hard.

More Related